Controlled experimental studies have repeatedly established that speakers adjust the referential expressions they use to take into account other candidate referents in a given context. These studies have been used to probe the extent to which reference strategies are communicatively “efficient” and indeed whether the efficiency of an expression is best defined in terms of the minimal number of entailments needed to differentiate a target referent, or in other terms, e.g., the speed with which an expression helps interlocutors coordinate on the target. Most of these studies have used decontextualized images which have, moreover, focused on very specific sets of potentially discriminating features, such as color or size. We present an analysis of data collected in a task that involved distinguishing target from distractor referents in naturalistic images; our goal was both to test the ecological validity of previous experimental studies as well as to identify additional discriminating features that had not been explored in previous research. Our results provide new evidence for the salience of entity parts and reveal finer-grained information about the relative uses of and dependencies between visual and other features, which can inform future experimental studies of how speakers communicate their targets of reference.